Professional Documents
Culture Documents
4 ISSN: 1837-7823
A Probabilistic Graph Based Approach for Interactive Image Segmentation and Object Extraction
Ashish Kumar M.Tech (CSE) LNCT BHOPAL INDIA asingh087@gmail.com
1
Prof. Uday Pratap Singh CSE Department LNCT BHOPAL, INDIA usinghiitg@gmail.com
Abstract
We propose a Probabilistic Graph based Model for interactive image segmentation. A multilayer probabilistic Graph is constructed from an oversegmentation of the image to model the relationships among super pixel regions, edge segments and vertices. We used an iterative procedure to merge several regions based on the probability of the regions. Regions are merged until the user is satisfied with the segmentation. Each nodes probability is updated after each iteration. With the help of proper user intervention the input image is segmented in short time period. We evaluate the proposed model on many images in the database. The results demonstrate that the Graphical model can be used for interactive image segmentation (IS). The results also shows that the proposed method has good accuracy and efficiency for both segmenting the image and extraction of the object from the segmented image.
Index Terms- Bayesian Network (BN), Graphical Model, interactive image segmentation (IS), image segmentation, region merging.
1. INTRODUCTION
Image segmentation is an important subject in the field of image processing and computer vision. It still remains as a challenge when the image involves complex natural scenes. Great efforts have been made to develop more advanced techniques in recent years. The process of partitioning a digital image into multiple regions or sets of homogeneous pixels is called image segmentation. Actually, partitions are different objects in image which have the same texture or colour. The result of image segmentation is a set of regions that collectively cover the entire image, or a set of contours extracted from the image. All the pixels in a region are similar with respect to some characteristics or computed property, such as colour, intensity, or texture. Adjacent regions are significantly different with respect to the same characteristics. This technique has a variety of applications including computer vision, image analysis, medical image processing, remote sensing and geographical information system. Image segmentation is based on two basic properties, first intensity values involving discontinuity that means sudden or abrupt changes in intensity as edges and second similarity that means partitioning a digital image into regions according to some pre-defined likeness criterion. Many approaches have been proposed earlier which includes region growing [1], normalized cuts [2], clustering [3], grab-cut method [4], active contours [5], and MRF (Markov Random Field) based approaches [6] etc. All these approaches are data driven approaches. The data driven approaches may not produce satisfactory results in segmentation when there are shadow, low contrast areas, occlusion, cluttering and noise present in the image. When these things are present in the image than prior knowledge is very necessary for segmenting the given image for producing satisfactory results. These methods incorporate the prior information in a deterministic manner, ignoring various uncertainties associated in the image segmentation process and making very difficult to take users intervention into the segmentation process. There are two basic types of graphical models: the undirected graphical model and the directed acyclic graphical model. The undirected graphical model can represent noncasual relationships among the random variables [7]. MRF models have been widely used for image segmentation. They incorporate the spatial relationships among neighbouring labels as a Markovian prior. This prior can encourage the adjacent pixels to be classified into the same group. As an extension to MRFs, the Conditional Random Field (CRF) [8] is another type of undirected graphical model that has become increasingly popular. While both MRF and CRF
International Journal of Computational Intelligence and Information Security, April 2012 Vol. 3, No. 4 ISSN: 1837-7823 models can effectively capture noncasual relationships they cannot model the directed relationships. This problem can be solved by another type of graphical model, i.e., the directed acyclic graphical model such as Bayesian Network (BN) [9] [10]. The automatic segmentation approaches may also fail even though they use prior information in segmentation process. The main reason behind this is the complexity of segmentation of image in real applications. The complexity is due to the several reasons like shadow, low contrast areas, occlusion, cluttering and noise in the image. These reasons make the segmentation process quite difficult and challenging. The use of interactive image segmentation process is the solutions of such problems. In the interactive image segmentation the proper users intervention is required for segmentation process. Because user gives the clue for segmenting the image in the process the results are more improved and satisfactory. We develop a graphical model for segmentation that can incorporate various probabilistic relationships and apply it to the segmentation problem. Our model captures the natural causal relationships among three entities in image segmentation: regions, edges, and vertices (i.e., the junctions). There are many works on interactive image segmentation [4],[11], [12, 13, 14]. They demonstrate the usefulness of the users intervention for improving segmentation. However, these approaches mainly exploit a limited type of user interventions. The user typically gives hard constraints by fixing the labels of certain pixels or group of pixels, without considering the uncertainties in the users intervention. Other types of users interventions are rarely used in these works.
2. RELATED WORK
Eric N. Mortensen and Jin Jia [15] proposed a BN model with two layers for image segmentation, which captures the relationships between edge segments and their vertices. Given a user input seed path, they use minimum path spanning tree (MPST) graph search to find the most likely object boundaries. They also encode statistical similarity measure between the adjacent regions of an edge into its a priori probability therefore implicitly integrating region information. In the early study [16] they use the similar BN model for both automatic and interactive segmentation. Their approach can find multiple non-overlapping closed contours before any given users intervention. The intervention will serve as an evidence to help select a single closed contour that covers the object of interest. Lei Zhang, Zhi Zeng, and Qiang Ji [17] proposed a method to extend the Chain Graph (CG) model. CG is a hybrid Probabilistic Graphical Model (PGM) which Contains both directed and undirected links. Its representation is powerful enough to capture heterogeneous relationships among image entities. For CG they first oversegment the image into superpixels and find out different heterogeneous relationships among image entities (superpixels, vertices or junctions, edges, regions etc.) They construct the CG model with parameterized links with derived Joint Probability Distribution (JPD). These links are represented by either potential function or conditional probabilities. They first create a Directed Master Graph then create undirected sub-graphs for some terms in the JPD of Directed Master Graph. They segment the image into two parts foreground and background. In the end they apply probabilistic inference in the foreground to find out most probable explanation. Kittipat Kampa, Duangmanee Putthividhya and Jose C. Principe [18] design a probabilistic unsupervised method called Irregular Tree Structure Bayesian Network (ITSBN). The ITSBN is constructed according to the similarity of image regions in an input image. ITSBN is a Directed acyclic graph (DAG) with two disjoint sets of random variables hidden and observed. The original image is oversegmented in multiscale hierarchical manner then they extract features from the input image corresponding to each superpixel. According to these superpixels ITSBN is built for each level. After applying the learning and inference algorithms the segmented image is produced. Fei Liu, Dongxiang Xu, Chun Yuan and William Kerwin [19] combined the BN (Bayesian Network) and MRF (Markov Random Field) to form an image segmentation approach. The BN generates a probability map for each pixel in the image and then MRF prior is incorporated to produce the segmentation. It is an interactive image
International Journal of Computational Intelligence and Information Security, April 2012 Vol. 3, No. 4 ISSN: 1837-7823 segmentation method. First each pixel will be individually assigned a probability value to be each given class. According to such probability map, BN provides a mechanism to convert the problem from feature space to image domain and they consider the prior knowledge on the image model and the spatial relationships between pixels, they use MRF based model to produce the segmentation results. Costas Panagiotakis, Ilias Grinias, and Georgios Tziritas [20] proposed a framework for image segmentation which uses feature extraction and clustering in the feature space followed by flooding and region merging techniques in the spatial domain, based on the computed features of classes. They use a new block-based unsupervised clustering method which ensures spatial coherence using an efficient hierarchical tree equipartition algorithm. They divide the image into different-different blocks based on the feature description computation. The image is partitioned using minimum spanning tree relationship and mallows distance. Then they apply Kcentroid clustering algorithm and Bhattacharya distance and compute the posteriori distributions and distances and perform initial labelling. Priority multiclass flooding algorithm is applied and in the end regions are merged so that segmentation results are produced. Li Zhang and Qiang Ji [21] have proposed a Bayesian Network (BN) Model for Both Automatic (Unsupervised) and Interactive (Supervised) image segmentation. They Construct a Multilayer BN from the oversegmentation of an image, which find object boundaries according to the measurements of regions, edges and vertices formed in the oversegmentation and model the relationships among the superpixel regions, edge segment, vertices, angles and their measurements. For Automatic Image Segmentation after the construction of BN model and belief propagation segmented image is produced. For Interactive Image Segmentation if segmentation results are not satisfactory then active input selection by user intervention selections are again carried out for segmentation. Lei Zhang, and Qiang Ji [22] develop a unified graphical model in which they combined directed graphical model and undirected graphical model. The combination allows capturing more complex and heterogeneous relationships among image entities. The unified model is more expressive and more powerful. But it only used for automatic segmentation not for interactive segmentation. They first propose to employ Conditional Random Field (CRF) to model the spatial relationships among image superpixel regions and their measurements. Then they introduce a multilayer Bayesian Network (BN) to model the causal dependencies that naturally exist among different image entities, including image regions, edges, and vertices. The CRF model and the BN model are then systematically and seamlessly combined through the theories of Factor Graph to form a unified probabilistic graphical model that captures the complex relationships among different image entities. Using the unified graphical model, image segmentation can be performed through a principled probabilistic inference.
3. PROPOSED FRAMEWORK
For interactive image segmentation, we need a model that can conveniently take the users intervention and according to the user intervention it produces the segmentation result. We propose a Graphical Model based on the regions (superpixels), edges, vertices (junctions) and the relationships between these image entities for interactive image segmentation. We start from the top left corner of the image and take the first pixel from each region. The pixel represents that whole region. Now we find out the connectivity of regions to each other. We take the first pixel and check for the connected regions and the first pixels of those regions. Now we connect all the pixels to each other. The flowchart of the proposed algorithm is given below. Given the edge map of an oversegmented image, we construct a graphical model (GM) to capture the relationships between regions, edge segments and vertices. Each pixel has the label that represents the region number to which the pixel belongs. The label graph is constructed when all the pixels are connected to each other. Here we calculate the probability of each region for merging with the other regions on the basis of colour histogram.
International Journal of Computational Intelligence and Information Security, April 2012 Vol. 3, No. 4 ISSN: 1837-7823
Input image
Initial segmentation
Calculate Probability
User Intervention
Region merging
NO
Extraction of object
We use an iterative and interactive approach for the segmentation of the image. User start the process and the model starts merging the regions, at the same time graph is generated to show the connectivity of the regions and the probability of regions to merging with the other regions. After first iteration some regions that are most probable merged with each other and results a new graph with less regions and fewer pixels and connectivity. Probability is calculated for each iteration. The calculated probability is the probability of the pixel to merge with other pixels in the graph. This process continues until the user is satisfied or there are no region remains in the image. Once the user is satisfied it can stop the process. The final segmentation result is obtained by the user intervention. The user can also interact with the final segmented image to extract the object of interest from the image.
International Journal of Computational Intelligence and Information Security, April 2012 Vol. 3, No. 4 ISSN: 1837-7823
If I is set of all image pixels, then by applying segmentation we get different unique regions like { R1, R2, R3,, Rn } which when combined formed I . Basic formulation is as follows: (a) n Ri = I where Ri Rj = i=1,n (b) Ri is a connected region, i=1, 2.n. (d) P(Ri Rj ) = FALSE for i j. (c) P(Ri) = TRUE for i=1, 2.. n.
Where P(Ri) is a logical predicate defined over the points in set Ri. Condition (a) indicates that segmentation must be complete; every pixel in the image must be covered by segmented regions. Segmented regions must be disjoint. Condition (b) requires that points in a region be connected in some predefined sense like 4-neighbourhood or 8-neighbourhood connectivity. Condition (c) deals, the properties must be satisfied by the pixels in a segmented region e.g. P(Ri) = TRUE if all pixels in Ri have the same gray level. Last condition (d) indicates that adjacent regions Ri and Rj are different in the sense of predicate P.
R1,P1
R2,P2
R3,P3
R4,P4
(a) (b)
R5,P5
R6,P6
International Journal of Computational Intelligence and Information Security, April 2012 Vol. 3, No. 4 ISSN: 1837-7823
P( R |R ,..., R
i 1 i =1
i 1
Thus, any region can be represented as a product of conditionals, e.g., P(R1,R2,R3) = P(R3 |R1,R2)P(R2,R1) = P(R3 |R1,R2)P(R2 |R1)P(R1)
Fig.4. Regions in the Image Here we choose to use the Bhattacharyya coefficient [25, 26, 27] to measure the similarity between R1 and R2.
( R1, R 2) =
u =1
4096
Hist .Hist
R1
u R2
Where HistR1 and HistR2 are the normalized histograms of R1 and R2, respectively, and the superscript u represents the uth element of them. Here is the Bhattacharyya coefficient. The higher the Bhattacharyya coefficient between R1 and R2 , the higher the similarity between them.
International Journal of Computational Intelligence and Information Security, April 2012 Vol. 3, No. 4 ISSN: 1837-7823
Fig. 5.
10
International Journal of Computational Intelligence and Information Security, April 2012 Vol. 3, No. 4 ISSN: 1837-7823
(f) Result
(g) Probablistic graph of (c) 2nd stage of merging Fig.6. Color image segmentation results arranged in three rows.
11
International Journal of Computational Intelligence and Information Security, April 2012 Vol. 3, No. 4 ISSN: 1837-7823
Fig.7. Examples of the colour image segmentation results arranged in groups of four rows. In each group, the first image is the colour image. The second image is the image mask and the third image is the segmentation result produced by the proposed approach. Our method is very simple as compared to the other existing methods of segmentation. The method is less time consuming and produced results quickly. Because it is interactive method so the time taken in the segmentation is totally depend on the user and input image. The segmentation speed mainly depends on the complexity of the constructed graphical model. Object extraction time is totally depending on the size of the object. We achieved encouraging results on these images. As to the images with small changes or similar colour of foreground and background, our algorithm will fail to achieve an ideal segmentation effect. Another kind of error is caused by the clutter. When the background (e.g., the shadow) has a similar appearance as the foreground, the proposed model may not be able to completely separate them.
6. CONCLUSION
To summarize, we present a new interactive image segmentation framework based on a probabilistic graphical model. The proposed graphical model can systematically capture the relationships among different image regions to perform effective image segmentation. An image is first oversegmented to produce an edge map, from which a superpixel-based Graphical Model is constructed. Probability is calculated for each node present in the graph. The model performs region-merging based on the colour histogram of the image. After region merging the new graph is constructed and new probability is calculated for each node in the graph. Maximum probability of a node shows that the region will be merged to other smaller regions present in the image and results in a bigger region. It is an iterative procedure and number of iterations depends on the user satisfaction. Finally, we want to point out that the application of the graphical model is not limited to image segmentation. It can find applications in many different computer vision problems including object tracking, object recognition etc. Our experimental results demonstrate the promising capability of the proposed Graphical model for effective interactive image segmentation. 12
International Journal of Computational Intelligence and Information Security, April 2012 Vol. 3, No. 4 ISSN: 1837-7823
REFERENCES
[1] L. Vincent and P. Soille, Watersheds in digital spaces: An efficient algorithm based on immersion
simulations, IEEE Trans. Pattern Anal. vol. 13, no. 6, pp. 583598, Jun 1991.
[2] J. Shi and J. Malik, Normalized cuts and image segmentation, IEEE Trans. Pattern Anal. Mach.
Trans. Pattern Anal. Mach. Intell., vol. 24, no. 5, pp. 603619, May 2002.
[4] Y. Boykov and M.-P. Jolly, Interactive graph cuts for optimal boundary and region segmentation of
objects in N-D images, in Int.Conf. Comput. Vis., 2001, vol. 1, pp. 105112.
[5] V. Caselles, R. Kimmel, and G. Sapiro, Geodesic active contours, Int. J. Comput. Vis., vol. 22, no. 1,
images, IEEE Trans. Pattern Anal. Mach. Intell., vol. PAMI-6, no. 6, no. 6, pp. 721741, Nov. 1984.
[7] L. Zhang and Q. Ji, Image segmentation with a unified graphical model, IEEE Trans. Pattern Anal.
Segmenting and Labeling Sequence Data, Proc. Intl Conf. Machine Learning, pp. 282-289,2001.
[9] F.V. Jensen, Bayesian Networks and decision Graph, Springer-Verlag, 2001. [10] J. Peral, Probabilistic Reasoning in Intelligent System: Network of Plausible Inference, Morgan-
3, pp. 303-308.
[14] E. N. Mortensen and W. A. Barrett, Intelligent scissors for image composition, in Proc. SIGGRAPH,
Proc. IEEE Conf. Comput. Vis. Pattern Recognit., 2006, pp. 1007-1014.
[16] E. Mortensen and J. Jia, A Bayesian network framework for real-time object selection, in Workshop
Human Activity Recognition and Image Segmentation, IEEE Transaction on Image Processing, VOL. 20, NO. 9, September 2011.
[18] Kittipat, Kampa, Duangmanee Putthividhya and Jose C. Principe, Irregular Tree-Structured Bayesian
Network for Image Segmentation., IEEE International Workshop on Machine Learning for Signal Processing, September 18-21, 2011.
13
International Journal of Computational Intelligence and Information Security, April 2012 Vol. 3, No. 4 ISSN: 1837-7823
[19] F. Liu, D. Xu, C. Yuan, and W. Kerwin, Image segmentation based on Bayesian network-Markov
random field model and its application on in vivo plaque composition, in Int. Symp. Biomed. Imag., 2006, pp. 141-144.
[20] Costas Panagiotakis, Ilias Grinias, and Georgeios Tziritas Natural Image Segmentaion Based on Tree
Equipartition, Bayesian Flooding and Region Merging, IEEE Transactions on Image Processing, Vol. 20, No. 8, August 2011.
[21] Lei Zhang and Qiang Ji, A Bayesian Network Model for Automatic and Interactive Image
Segmentation, IEEE Transaction on Image Processing, VOL. 20, NO. 9, September 2011.
[22] Lei Zhang and Qiang Ji, Image Segmentation with a Unified graphical model, IEEE Transaction on
Pattern Analysis and Machine Intelligence, VOL. 32, NO. 8, AUGUST 2010.
[23] S. Birchfield, Elliptical head tracking using intensity gradients and color histograms, in: Proceedings
of IEEE Conference on Computer Vision and Pattern Recognition, 1998, pp. 232-237.
[24] T.Ojala,
M.Pietikainen,
P.Maenpaa
Multiresolutiongray-scaleandrotation
invariant
texture
classification with local binary patterns, IEEE Transactionson Pattern Analysis and Machine Intelligence, 2002, pp.971-987.
[25] M.J. Swain, D.H. Ballard, Color indexing, International Journal of Computer Vision Vol. 7 No. 1,
14